Add ImageNet pre-trained weights under `pretrained` feature flag

[laggui]

This PR adds new methods to directly initialize a ResNet-{18, 34, 50, 101, 152} with ImageNet pre-trained weights from torchvision.

The weights are automatically downloaded from the web to a default ~/.cache/resnet-burn/ folder using download_file_as_bytes which provides a progress bar.

Changes:

• Refactor all modules to have configs with init and init_with methods following good practice
• Add pretrained feature flag
  • New weights module
  • New resnet*_pretrained methods
• Refactor example accordingly

Because loading the pre-trained weights requires a fix not yet in a released version of candle-core, the current burn dependency is pinned to a specific revision that pins the correct dependency.

TODO:

• [ ] Change burn dependency to a specific version once the new candle-core version is released and included in a burn release/patch

[laggui]

@antimora tagging you specifically regarding the pattern we might want to adopt for pre-trained weights as discussed in #16

[antimora]

Thank you! I'll review it soon.

[laggui]

Also fixed some minor issues with the bigger ResNet models.